ICCV19-Paper-Review

Summaries of ICCV 2019 papers.

Goal-Driven Sequential Data Abstraction

Benchmarking machine intelligence and supporting summarization applications importantly requires automatic data abstraction. The former deals with the apprehension of input data to produce a meaningful but more compact abstraction. Whereas, the latter exploits this capability to save space or human time by summarizing the essence of input data.

This paper deals with the study of a general re-inforcement learning based framework for learning to abstract sequential data in a goal-driven way.

Model Features

The above model namely GDSA(Goal Driven Sequence Abstraction) is driven by a goal-function rather than needing expensively annotated ground-truth labels, and also uniquely allows selection of the information to be preserved rather than producing a single general-purpose summary. It provides improved performance for abstraction compared to several alternatives.

GDSA v/s other Models

Reduced data requirements and new goal-conditional abstraction ability enables unique practical summarization applications.

Tensorflow implementation of GDSA for Sketch Abstraction can be found here.